DeepPrecip: a deep neural network for precipitation retrievals

Abstract. Remotely-sensed precipitation retrievals are critical for advancing our understanding of global energy and hydrologic cycles in remote regions. Radar reflectivity profiles the lower atmosphere commonly linked to through empirical power laws, but these relationships tightly coupled particle microphysical assumptions that do not generalize well different regional climates. Here, we develop a robust, highly generalized retrieval algorithm from deep convolutional neural network (DeepPrecip) estimate 20 min average surface accumulation using near-surface radar data inputs. DeepPrecip displays high skill can accurately model total accumulation, with mean square error (MSE) 160 % lower, on average, than current methods. also outperforms less complex machine learning algorithm, demonstrating value when applied retrievals. Predictor importance analyses suggest combination both (below 1 km) higher-altitude (1.5–2 measurements primary features contributing accuracy. Further, closely captures magnitudes variability across nine distinct locations without requiring any explicit descriptions microphysics or geospatial covariates. This research reveals important role extracting relevant information about atmospheric